2

I would like to get some tips, because I am currently stuck with my regression part. I am using binary logistic regression. The aim is to compare two separate groups' responses. I have a MV (moderator variable, 1-one group, 0-other group). My IV (independent variables are 11 and all of them were measured on 7-point Likert scale). My DV (dependent variable or outcome is choice of 2 categories). First problem I face: I used the median split approach to obtain my DV for each respondent, as I do not have my DV, which I have to obtain to run regression. Second, I tried to run 2 separate regressions for two groups, though my coefficients were poorly significant. Than I tried to run a one model with interactions. Calculating product term for each IV (IV1mean_centered* MVmean_centered) etc. for all IV I have. While running a regression I entered all the IV's vs. my MV and also calculated product terms with MV_centered (also tried without mean centering and with only MV).

My issue is that my coefficients are even less significant than with separate models!!! Does anyone have some suggestions how to solve this issue? Maybe I have made some mistake regarding interactions in logistic regression?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
user18213
  • 21
  • 1
  • 5
    Your post isn't too clear. Perhaps you could write out the equations you're estimating? Also, this is the first I've seen of "median splitting" -- but if you're arbitrarily dividing a continuous variable and creating a binary variable, you might imagine that you are muddying your data. Perhaps drop those in the 40-60 percentile? Finally, it is possible that your coefficients are not significant because they are either collinear, or they arenit in fact good predictors. – generic_user Dec 31 '12 at 13:32
  • 1
    I agree w/ @ACD, this post isn't clear enough to answer. One thing I can say is that you shouldn't perform median splits, or categorize your data in other ways. For reference, I wrote about that here: [how to choose between ANOVA & ANCOVA in a designed experiment](http://stats.stackexchange.com/questions/24077//24080#24080). – gung - Reinstate Monica Dec 31 '12 at 13:41
  • 1
    This sounds like you are making a mistake. You shouldn't have to do things to get your dependent variable - that should be part of the data, although you might have to transform it – Peter Flom Dec 31 '12 at 20:00
  • Also, it might not be right to treat your Likert scale variables that way - it's very often done, and often causes no problems, but it might be problematic here – Peter Flom Dec 31 '12 at 20:01

0 Answers0